Although we usually talk about NVIDIA as a company dedicated to gaming hardware, the company is much more. Above all, it is very well valued within the advanced computing and Artificial Intelligence market. To meet the needs in these two fields, the company has just announced its NVIDIA A100 PCIe card. This new accelerator promises great performance with a tight consumption.
This new NVIDIA A100 PCIe, as the company states, is designed for “AI, data science, and scientific computing.” It is also characterized by being based on the Ampere architecture and therefore uses a 7nm manufacturing process. Additionally, this graphics card is based on PCI-Express 4.0, so it is designed to work with AMD EPYC processors.
[amazon box="B07MJGCPW5"]NVIDIA A100 PCIe, a high power and low consumption accelerator
The new NVIDIA A100 PCIe in specifications is very similar to the NVIDIA A100 SXM model, but with a reduced TDP. It goes from a TDP of 400W for the SXM model to a TDP of 250W for this A100 PCIe model.
This NVIDIA solution features an 826mm² die with a staggering 54 billion transistors. It integrates 6912 CUDA cores and 432 Tensor cores at a frequency of 1.410MHz. It doesn't utilize a single silicon die with all CUDA cores, so we could see even more powerful versions. This solution boasts 19.5 TFLOPs of FP32 performance, 9.7 TFLOPs of FP64 performance, and 19.5 TFLOPs of FP4 Tensor core performance.
Additionally, this NVIDIA A100 PCIe has 40GB HBM2E with a memory interface of 5.120bits with a bandwidth of 1.555GB / s. This solution makes use of a third generation NVLink system for a bandwidth of 4.8TB / s. Additionally, this solution is based on the PCI-Express 4.0 / SXM4 interface.
Because it's designed for Artificial Intelligence , it features new Tensor Cores that are twice as fast as the previous generation. Additionally, it can be up to 20 times faster than the Tesla V100 for AI tasks.
Press release
NVIDIA and the world's leading server manufacturers have just introduced their systems based on NVIDIA A100 technology with a wide variety of designs and configurations to meet the most complex challenges of AI, data science and scientific computing.
Following the release of the NVIDIA Ampere architecture and NVIDIA A100 GPU last month, more than 50 A100 servers from the world's leading manufacturers are expected to be installed, including Asus, Atos, Cisco, Dell Technologies, Fujitsu. , Gigabyte, Hewlett Packard Enterprise, Inspur, Lenovo, One Stop Systems, Quanta / QCT, and Supermicro.
Server availability varies, with 30 systems planned this summer, and more than 20 more by the end of the year.
“Adoption of NVIDIA A100 GPUs across leading server vendors’ offerings is outpacing anything we’ve seen before,” said Ian Buck, vice president and general manager of accelerated computing at NVIDIA. “The sheer breadth of NVIDIA A100 servers from our partners ensures that customers can choose the best options to accelerate their data centers for high utilization and low total cost of ownership.”
The first GPU based on NVIDIA's Ampere architecture, the A100 can boost performance up to 20 times more than its predecessor, making it the company's biggest GPU performance leap to date. It includes several technical advancements, including a new multi-instance GPU technology that allows a single A100 to be split into up to seven different GPUs to handle different computing tasks; XNUMXrd generation NVIDIA NVLink technology that allows multiple GPUs to be joined together to function as a single giant GPU; and new structural spreading features that can be used to double the performance of a GPU.
